Skip to content
motiondetection

Person detection camera software, no new camera required

Buying a security camera with person detection means buying a camera. Adding person detection to the cameras on your wall means changing the software that reads them, and the second one is cheaper, faster and reversible.
Run it on a clip
  • Person, not identity
  • Works under infrared
  • No face recognition, ever

Interior, after hours

Fixed camera, low light, two subjects

00:00.0

Clip

Decoded in your browser. Your video is never uploaded.

6

Sets the luma delta threshold and the smallest region that counts.

0.4s

Kills the one frame flicker that no operator wants paged about.

Exclude zone

Put it over the road, the tree, the neighbour's window. Events inside it stop alerting.

The frame an event was found on shows up here.

00 00:00.0

Nothing selected yet

Hover a row in the event table to preview its frame, click it to hold the frame and jump the clip there.

Length
0.0s
Confidence
0%
Area
0.0%

0

Motion events found

0

Dropped by zone

0

Dropped, too short

0

Alerts you would get

# Start Length Class Conf. Area

Reading the clip. Events appear here as they are found.

01 Person detection is not face recognition

These get confused constantly, and the difference is the entire legal question.

If you take one thing from this page to your counsel, take this section.

What this does

Person detection

The model answers one question about a cropped frame: what category of object is this? One answer is person. It does not know, and cannot know, which person. Nothing is stored that could be compared against another frame to say the same individual appeared twice.

  • Answers: a person was in this area, at this time, for this long.
  • Stores: a class, a confidence value, a timestamp, a duration, a motion area.
  • Computes no biometric identifier of any kind.

What this refuses to do

Face recognition

Face recognition computes a mathematical template from a face and compares it against stored templates to name an individual. That template is a biometric identifier, and in Illinois, Texas and Washington collecting one carries specific statutory obligations, in some cases with a private right of action.

  • We build no face templates, ever.
  • We do no gait analysis and no re identification between cameras or between days.
  • This is a permanent product decision, not a feature we have not reached yet.

We are not your lawyers and this is not legal advice. The fact that matters for the analysis is the one stated above, and it is written in questionnaire language on the enterprise security camera systems page.

02 Why person is the class that matters

Almost every site ends up alerting on one class and reading the rest as a list.

After a week of real use, most operators converge on the same configuration, and it looks like this.

Class Typical routing Why
person Phone, all hours The only class where a false negative is expensive. Nearly every incident that matters has a person in it somewhere.
vehicle Phone, outside opening hours During the day a vehicle is a customer. At 02:00 on a closed site it is the thing that brought the person.
package Timeline, or phone at a door Useful at entrances and mail rooms, noise everywhere else.
animal Timeline only The second class most people mute. Worth keeping in the record, not worth a notification at 03:00.
other motion Timeline only Weather, flexing doors, flags, light changes. The bucket that used to be your entire alert channel.

Nothing is deleted by this configuration. Every class stays in the timeline with its timestamp, so the morning after an incident you still have the vehicle that arrived at 01:47, even though it never rang a phone.

03 Against the person detection in a new camera

Your camera's own person filter is genuinely useful, and here is exactly where it stops.

Condition On camera filter This
Good light, close range Works well Works well
Distance, past about twenty metres Often misses Depends on pixels on target, and the confidence value says so
Under infrared at night Frequently confused by bright close objects Classifies the crop, so a moth is not a person
A road or footpath in frame Rectangular regions only Polygon exclude zones
Short flicker events Usually no duration control Minimum duration from 0.2 to 5 seconds
Across a mixed brand estate Only on the newer cameras Every camera that publishes a stream, whatever its age

If your cameras have an on device person filter, turn it on first. It is free, it is already paid for, and if it gets you quiet nights you do not need this product. What it cannot give you is polygon zones, duration control and one consistent behaviour across an estate that was bought in three different decades.

What decides accuracy, in order

  1. Pixels on target. A person needs to be enough of the frame to have a shape. This is a lens and placement question before it is a model question.
  2. Camera angle. Across the space, not down it. A near overhead view is the hardest for any model.
  3. Illumination placement. An infrared lamp mounted away from the lens transforms night results.
  4. A clean lens. Free, and it beats every setting on this page.
  5. Then the model. Which is fifth, and anybody selling you the model as first is selling.

The physics behind items two to four is in why security cameras send false triggers.

04 Asked about person detection

Four questions before you turn it on.

01

Can it tell staff from strangers?

No, and it will not try. It reports that a person was present at a time, in an area, for a duration. Deciding whether that person had a reason to be there is your job, with your access log and your gate codes, and that boundary is deliberate.

02

Does it work in the dark?

It works on whatever the camera can see. Under infrared with reasonable illumination, yes. In genuine darkness with no infrared, nothing works, including a human being watching the same monitor.

03

Will it alert on a person inside a vehicle?

Usually the event is classified as vehicle, because that is what dominates the frame. If somebody gets out, that becomes its own person event a moment later, which is the sequence that actually matters on a closed site at night.

04

What happens to events it is unsure about?

They arrive with the confidence value the model returned, shown as it came back. If the call fails entirely, the event is labelled unclassified and keeps its real motion data. Nothing is ever guessed into a class.

05 Read next

Related pages on classes and alerts

Motion detection software

The detection stage that finds the events in the first place.

Artificial intelligence for video surveillance

What the model gets, returns and refuses.

RTSP software

Connecting the camera the person will walk past.

Self storage security cameras

Person alerts on a site made of moving metal doors.

Enterprise security camera systems

The biometrics answer, written for a questionnaire.

Video analytics software price

What person detection costs per camera per month.

Walk past your own camera, then read the row.

Record a clip, run it in the demo, and see whether the event says person with a confidence value you would trust at two in the morning.

Run it on a clip first